AI TI-RADS

AI TI-RADS refines thyroid nodule risk stratification by applying a genetic algorithm to recalibrate ACR TI-RADS feature point assignments using biopsy-proven nodules.

Key Features:

• Genetic algorithm optimization: Leverages a genetic algorithm applied to a dataset of 1325 biopsy-proven thyroid nodules from 1264 patients to recalibrate point values assigned to ACR TI-RADS features including composition, echogenicity, shape, margin, and echogenic foci.
• Simplified feature assignment: Reassigns zero points to six of the eight ACR TI-RADS features to streamline categorization.
• Diagnostic performance metrics: Demonstrates increased expert-reader specificity (65% vs 47% for ACR TI-RADS) and an area under the ROC curve of 0.93 versus 0.91 for ACR TI-RADS.
• Improved performance across readers: Shows increased mean specificity for nonexpert readers from 48% with ACR TI-RADS to 55% with AI TI-RADS.
• Reader-based evaluation: Performance was assessed relative to both expert and nonexpert radiologist interpretations.

Scientific Applications:

• Risk stratification: Provides more precise risk stratification of thyroid nodules to inform clinical decision-making.
• Biopsy reduction: Improved specificity supports reduction of unnecessary thyroid biopsies and associated invasive procedures.
• Mitigation of reader variability: Enhances diagnostic consistency across varying levels of radiological expertise.

Methodology:

Retrospective analysis (reported for 1425 biopsy-proven thyroid nodules) with expert readers assigning ACR TI-RADS points, optimization of point assignments using a genetic algorithm trained on a subset of nodules, and performance evaluation using binomial proportion tests and bootstrapping against expert and nonexpert interpretations.